Enhancing Pediatric Distal Radius Fracture Detection: Optimizing YOLOv8 with Advanced AI and Machine Learning Techniques.

doi:10.21203/rs.3.rs-5306607/v1

Background

In emergency departments, residents and physicians interpret X-rays to identify fractures, with distal radius fractures being the most common in children. Skilled radiologists typically ensure accurate readings in well-resourced hospitals, but rural areas often lack this expertise, leading to lower diagnostic accuracy and potential delays in treatment. Machine learning systems offer promising solutions by detecting subtle features that non-experts might miss. Recent advancements, including YOLOv8 and its attention-mechanism models, YOLOv8-AM, have shown potential in automated fracture detection. This study aims to refine the YOLOv8-AM model to improve the detection of distal radius fractures in pediatric patients by integrating targeted improvements and new attention mechanisms.

Methods

We enhanced the YOLOv8-AM model to improve pediatric wrist fracture detection, maintaining the YOLOv8 backbone while integrating attention mechanisms such as the Convolutional Block Attention Module (CBAM) and the Global Context (GC) block. We optimized the model through hyperparameter tuning, implementing data cleaning, augmentation, and normalization techniques using the GRAZPEDWRI-DX dataset. This process addressed class imbalances and significantly improved model performance, with mean Average Precision (mAP) increasing from 63.6% to 66.32%.

Results and Discussion

The iYOLOv8 models demonstrated substantial improvements in performance metrics. The iYOLOv8 + GC model achieved the highest precision at 97.2%, with an F1-score of 67% and an mAP50 of 69.5%, requiring only 3.62 hours of training time. In comparison, the iYOLOv8 + ECA model reached 96.7% precision, significantly reducing training time from 8.54 to 2.16 hours. The various iYOLOv8-AM models achieved an average accuracy of 96.42% in fracture detection, although performance for detecting bone anomalies and soft tissues was lower due to dataset constraints. The improvements highlight the model's effectiveness in pathological detection of the pediatric distal radius, suggesting that integrating these AI models into clinical practice could significantly enhance diagnostic efficiency.

Conclusion

Our improved YOLOv8-AM model, incorporating the GC attention mechanism, demonstrated superior speed and accuracy in pediatric distal radius fracture detection while reducing training time. Future research should explore additional features to further enhance detection capabilities in other musculoskeletal areas, as this model has the potential to adapt to various fracture types with appropriate training.

In hospital emergency departments (EDs), radiologists interpret numerous X-rays to detect fractures and abnormalities. Among pediatric populations, distal radius fractures (DRF) are the most common type of traumatic injury [1]. These X-rays are generally analyzed accurately in well-resourced, high-quality hospitals with experienced radiologists. However, inexperienced residents may be left to interpret X-rays in rural areas where radiologists are often absent. This situation can lead to lower diagnostic accuracy, potentially resulting in delayed or inadequate treatment, prolonged pain, and increased operative risks [2, 3].

Even in high-resource settings, radiologists are not immune to diagnostic errors. Recent reviews indicate that X-ray misinterpretation rates can reach 26%, with fractures being among the most commonly missed diagnoses and a frequent source of medical-legal claims [4, 5].

To mitigate these challenges, machine learning (ML) systems offer a promising solution. By analyzing large datasets of medical images, ML algorithms can identify subtle features, such as microfractures in the distal radius, that human interpreters might miss. This technology has the potential to significantly enhance diagnostic accuracy and improve patient care in both urban and rural ED settings. In recent years, artificial intelligence (AI) and deep learning have become invaluable assets in radiography and diagnostic medicine [6]. Automated systems have demonstrated significant efficiency in detecting conditions such as pneumonia in chest radiographs, diagnosing breast cancer, and assessing diabetic retinopathy, among other applications [7]. By building on these advancements, Researchers have begun applying AI systems and object detection models to fracture detection [8, 9]. The diagnostic reliability of these models is on par with that of medical experts, marking a revolutionary shift in medical imaging diagnosis and analysis [10].

In 2023, Ultralytics introduced the latest version of the YOLO (You Only Look Once) series, YOLOv8. Building upon this, Chien et al. developed YOLOv8-AM, which integrates attention mechanisms into the YOLOv8 original architecture. Studies have shown that incorporating attention mechanisms into convolutional blocks can significantly enhance performance [11–13]. However, YOLOv8-AM has faced limitations, including high training times and moderate precision.

Our research aims to address these challenges by refining the YOLOv8-AM model through targeted improvements and adding a new attention mechanism. We will evaluate the enhanced model's effectiveness in detecting DRF in the pediatric population, seeking to improve accuracy and robustness. Finally, to provide freely this detecting tool online.

In this work, we utilized and enhanced the existing methodologies proposed by Rui-Yang Ju et al. in their paper on pediatric wrist fracture detection using YOLOv8 and attention mechanisms [14, 15]. The architecture of our model follows a similar design as presented in their work, explicitly building upon the YOLOv8 backbone and integrating attention mechanisms to improve detection accuracy. We refer to the original architecture proposed by Rui-Yang Ju for detailed insights into the basic framework, and the improvements introduced in this work are elaborated below.

The YOLOv8 architecture serves as the foundation for this work. It consists of four key components: the Backbone, Neck, Head, and Loss Function, and is largely based on the structure proposed by Chien et al. [16]:

Backbone: The Cross-Stage Partial network forms the backbone, optimized for computational efficiency. YOLOv8 replaces YOLOv5's C3 module with the C2f module, enhancing feature extraction while reducing computational load. The Convolution-Batch Normalization-SiLU structure is used in all convolutional layers.
Neck: YOLOv8 combines Feature Pyramid Networks and Path Aggregation Networks for multi-scale feature extraction. Following Ju et al. [15], we made minor modifications, including attention modules.
Head: YOLOv8 adopts a decoupled head structure, allowing separate classification and regression processing. It uses an anchor-free approach, improving accuracy for small objects like fractures.
Loss Function: YOLOv8 uses Binary Cross-Entropy for classification and Distribute Focal Loss with Complete Intersection over Union for regression, enhancing small object detection.

Hyperparameter tuning were conducted to enhance our models' performance and develop the improved YOLOv8 (iYOLOv8) model. We began our experiments by training the model for 60 epochs, as recommended by baseline YOLOv8 studies. However, we quickly discovered that increasing epochs yielded better results. Systematically testing up to 100 epochs revealed significant improvements in precision and recall.

Curious about the potential benefits of extended training, experimentation with 300 epochs was also performed. While this did result in a slight increase in accuracy, we noted diminishing returns beyond 100 epochs, with only marginal improvements in mean Average Precision (mAP) and longer training times. The optimal learning rate was identified through several iterations as 1e^− 2, paired with a weight decay of 5e^− 4. This combination allowed the model to converge quickly without overfitting. Additionally, a batch size of 16 specifically for fracture detection in pediatric wrist X-rays was selected, striking a balance between computational efficiency and model performance. This size facilitates stable gradient updates while preserving the nuances of small-scale features critical for accurate fracture identification. The SGD optimizer was preferred over Adam due to its superior high-dimensional medical image data handling. Specifically, SGD demonstrated more consistent convergence in refining model weights for fracture detection tasks, ultimately enhancing feature extraction and classification accuracy for subtle fractures. The newly modified architecture of the model is illustrated in Fig. 1.

To further refine feature extraction and bolster the model's ability to identify fractures in pediatric wrist X-rays, multiple attention mechanisms (AM) were incorporated into the architecture, which led to iYOLOv8-AM models. These include the Convolutional Block Attention Module (CBAM), Global Attention Mechanism (GAM), Efficient Channel Attention (ECA), Shuffle Attention (SA), and Global Context (GC) Block Development (Fig. 2). Each of these modules was independently added after the four C2f modules in the Neck, enabling the model to selectively focus on the most relevant features while effectively suppressing irrelevant information.

CBAM: Sequentially applies Channel and Spatial Attention to emphasize informative parts of the image.
GAM: Simplifies feature recalibration, removing max pooling to preserve details in medical images better.
ECA: Utilizes 1D convolution for efficient channel-wise attention, improving feature integration.
SA: Uses Channel Shuffle to focus on grouped feature maps, balancing accuracy and efficiency.
GC Block: Captures both global and local features, which are crucial for identifying subtle wrist fractures.

One of the primary innovations in this research was developing and refining the GC block, which proved to be the most effective attention mechanism compared to others such as SA, ECA, and GAM. While the GC block had been previously introduced in object detection models, we proposed critical structural improvements to make it more powerful and efficient in medical image analysis, particularly for fracture detection (Fig. 3).

The original GC block was designed to capture global information from images, enhancing the network's ability to handle complex object detection tasks by aggregating global features. However, certain inefficiencies were identified in addressing more minor features, such as subtle fractures in medical images. To tackle these shortcomings, several modifications were proposed. In the original GC block, global and local features were aggregated without prioritizing critical regions within the image. To improve this, a dynamic weighting mechanism that assigns greater importance to regions likely to contain fractures while still considering the global context was introduced [17]. This adjustment allows the model to focus more on relevant areas, such as bone structures in X-rays, while filtering out irrelevant background noise.

Let the feature map be denoted as $\:F\:\in\:{R}^{C\times\:H\times\:W}$, where C is the number of channels, and H and W are the height and width of the feature map. Dynamic weighting is applied using a learned weighting map. $\:W\:\in\:{R}^{C\times\:H\times\:W}$, which modifies the feature map by element-wise multiplication:

$$\:{F}_{weighted\:}=F\:⨀\:W$$

In this case, $\:⨀$W represents element-wise multiplication, and it is generated through a learned function that applies more significant weight to regions with high fracture likelihood. This helps the model focus on relevant areas.

Moreover, the standard GC block utilized a static global pooling layer, often resulting in the loss of detailed spatial information crucial for fracture detection. To address this, we proposed an adaptive pooling layer that adjusts the pooling size based on the detected features. This adaptive pooling ensures that more minor features, such as fine fractures, are preserved during feature extraction while capturing the broader global context. Adaptive pooling is performed with varying sizes for an input feature map FFF to maintain international and local features. Let $\:{P}_{s}$(F) be the adaptive pooling operation with size s. The final output is a combination of pooled features at multiple scales:

$$\:{F}_{pooled}=Concat({P}_{1}\left(F\right),\:{P}_{2}\left(F\right),\:{P}_{3}\left(F\right),\dots\:)\:$$

Additionally, the GC block was enhanced with cross-dimensional interactions to improve the feature refinement process, allowing it to learn dependencies between spatial and channel dimensions more effectively [18]. This change enables the model to process spatial and contextual information jointly, improving the overall feature representation of both small and large fractures. For a feature map F, this is expressed as:

=(F) (F)

$\:{F}_{interaction}$ =$\:{F}_{c}$(F) $\:⨀$ $\:{F}_{s}$(F)

Where $\:{F}_{c}$(F) denotes the Channel attention map, $\:{F}_{s}$(F) denotes the spatial attention map, and $\:⨀$ denotes element-wise multiplication. The GC block's effectiveness was enhanced while also focusing on computational efficiency. By streamlining the feature aggregation process and reducing redundant operations, the GC block maintained a low inference time of 8.2 ms, critical for real-time medical applications.

Key parameters and metrics to define the models’ performance:

Epochs: Represents one full cycle where the model goes through the entire dataset during training. Each epoch helps the model learn and refine its internal parameters to improve accuracy in predicting fractures.
Parameters (PARMS): Internal values that the model learns during training. These include weights and biases, which are adjusted to minimize error and improve the fracture detection performance.
Inference: Phase where the trained model is used to make predictions on new data, such as detecting fractures in medical images after the model has been trained.
Precision: Proportion of correctly predicted positive cases (true positives) out of all predicted positive cases (both true positives and false positives). It tells us how reliable the positive predictions are.
Recall: Proportion of actual positive cases (true positives) that the model correctly predicted. It reflects the model's ability to detect all relevant cases.
F1-Score: Combines precision and recall into a single metric to assess the model’s overall accuracy, especially in cases where there’s an imbalance between the number of fracture and non-fracture instances.
mAP50 (Mean Average Precision at IoU 50%): Model's average detection accuracy when using a 50% overlap threshold between predicted bounding boxes and the actual location of fractures. It’s commonly used to evaluate object detection tasks like medical imaging.
mAP95 (Mean Average Precision at IoU 95%): mAP95 extends mAP50 by calculating the average precision across multiple IoU thresholds (ranging from 50–95%), providing a more comprehensive assessment of the model’s ability to locate fractures accurately.
FLOPs (Floating-Point Operations) Quantifies the computational complexity of the model by counting the number of floating-point operations needed during inference. It indicates how much computational effort is required to detect fractures in new data.

The GRAZPEDWRI-DX dataset was used, comprising over 20,000 X-ray images, to detect pediatric wrist fractures. To further enhance the model's performance, several steps were implemented to improve the dataset's quality.

First, the dataset underwent a thorough cleaning process, during which low-quality images—such as those with artifacts or poor resolution—were removed. Mislabeling issues were also addressed by cross-referencing image annotations with expert radiologist reviews, with particular attention to underrepresented cases like "bone anomaly" fractures.

Another significant challenge in the dataset was the imbalance between different fracture types. To mitigate this, synthetic data augmentation techniques was employed, including random rotations, flips, and brightness adjustments, specifically targeting minority classes such as "soft tissue" and "bone anomaly" fractures. This approach enhanced the model’s ability to detect these rare fractures.

Additionally, the brightness and contrast of the X-ray images were normalized to achieve greater uniformity across the dataset. This step reduced noise and allowed the model to generalize better across various X-ray sources. To ensure robust evaluation, stratified random split was done to create balanced training, validation, and test sets, preserving the ratio of different classes in each split. This strategy improved the model’s generalization capability and helped reduce overfitting.

Integrating these attention mechanisms and the dataset improvements resulted in substantial performance gains over the baseline YOLOv8 model. Specifically, the mAP50 improved from 63.6–66.32%, surpassing previous state-of-the-art results. Remarkably, the model maintained an efficient inference time, increasing by only 0.2 ms per image despite the added complexity. Moreover, detection accuracy was notably enhanced for challenging cases, such as small fractures and underrepresented classes, thanks to the attention mechanisms and the improved dataset balance.

Combining the new iYOLOv8 architecture with advanced attention mechanisms and dataset enhancements, this work offers a robust solution for pediatric wrist fracture detection, demonstrating significant improvements in accuracy and efficiency.

The iYOLOv8 models demonstrated significantly improved precision and other performance metrics compared to the original model. The iYOLOv8 + GC model achieved the highest precision in detecting fractures at 97.2%, with an F1-score of 67% and a mAP50 of 69.5%, requiring only 3.62 hours of training. This reflects increased performance in fracture detection due to the GC block (Table 1).

Table 1: Improved Performance Metrics of iYOLOv8_AM Variants in Pediatric Distal Radius Fracture Detection for an input image size of 640.

Model	Epochs	Precision (F)	Recall	F1-score	mAP50 [%]	mAP50 – 95[%]	Training Time (hours)	Inference Time (ms)
iYOLOv8	100	0.959	0.88	0.63	65.2	43.21	2.17	7.7
iYOLOv8 + SA	100	0.961	0.84	0.65	65.9	41.29	2.28	3.4
iYOLOv8 +ECA	100	0.967	0.87	0.66	66.1	43.9	2.16	8.0
iYOLOv8 + GAM	100	0.957	0.88	0.63	66.1	44.2	2.37	7.1
iYOLOv8 + ResBlockCBAM	100	0.969	0.87	0.64	67.2	44.5	2.4	2.3
iYOLOv8 + GC	100	0.972	0.93	0.67	69.5	45.1	3.62	8.2

The iYOLOv8 + ECA model achieved a precision of 96.7% compared to 89.8% in the original model and an F1-score of 66% versus 64%. Although the mAP50 remained relatively stable, the model benefited from a significant reduction in training time from 8.54 to 2.16 hours (Tables 1 & 2). This suggests that ECA effectively enhances detection accuracy. The iYOLOv8 + ResBlockCBAM model managed a balanced improvement in precision to 96.9% while improving mAP50 to 67.2% and significantly reducing both training and inference times. However, incorporating advanced AM, especially the GC block, increased training times due to higher computational demands than models without additional AMs. Despite this, the overall training time for the improved models decreased compared to the original, and inference times remained relatively stable across all models. To note, the ResBlockCBAM variant achieved the lowest inference time of 2.3 ms (Tables 1 & 2).

Table 2: Original Performance Metrics of YOLOv8_AM Variants in Pediatric Distal Radius Fracture Detection for an input image size of 640.

Model	Epochs	Precision (F)	Recall	F1-score	mAP50 [%]	mAP50 – 95[%]	Training Time (hours)	Inference Time (ms)
YOLOv8	60	0.909	0.889	0.63	63.6	40.4	8.48	7.7
YOLOv8 + SA	60	0.868	0.89	0.62	65.8	42.2	3.37	8.7
YOLOv8 +ECA	60	0.898	0.899	0.64	64.3	41.6	8.54	8.0
YOLOv8 + GAM	60	0.851	0.898	0.61	64.2	41.9	4.868	7.7
YOLOv8 + ResBlockCBAM	60	0.893	0.881	0.64	64.2	41.0	3.608	12.7

As demonstrated in Table 3, adding AM to the iYOLOv8 model enhances overall precision, with the iYOLOv8_GC model achieving the highest precision at 75.2%. All models feature parameters comparable to those of medium and large models for an input size of 640, as outlined in the Chien et al. study, indicating that the models deliver high performance while being relatively average to large size. Our FLOPs are like those of large model sizes, suggesting that the computational resources used to develop and evaluate our models are robust and highly efficient.

Table 3: Improved Performance Metrics of iYOLOv8_AM Variants in All Categories Detection in Pediatric Distal Radius for an input image size of 640.

Model	Precision	Recall	F1-score	mAP50 [%]	mAP50 – 95[%]	Params	FLOPs	Inference Time (ms)
iYOLOv8	0.689	0.701	0.632	0.652	0.432	43.61M	165.3M	7.7
iYOLOv8 + SA	0.691	0.723	0.654	0.659	0.412	53.87M	199.1M	3.4
iYOLOv8 +ECA	0.731	0.794	0.661	0.661	0.439	43.64M	172.3M	8.0
iYOLOv8 + GAM	0.725	0.762	0.632	0.661	0.442	43.64M	166.8M	7.1
iYOLOv8 + ResBlockCBAM	0.672	0.731	0.641	0.672	0.445	49.29M	186.9M	2.3
iYOLOv8 + GC	0.752	0.801	0.678	0.695	0.451	49.29M	207.3M	8.2

From Figure 4, it is evident that the various iYOLOv8-AM models excel in detecting fractures, metal, and text, achieving an average accuracy of 96.42% compared to 88.4% in the original models for fracture detection, with the GC model attaining the highest precision and recall. Notably, as precision improves across all models, recall remains high until it reaches 80%, after which precision decreases as recall surpasses 80%. Conversely, the models demonstrate lower accuracy in detecting bone anomalies, bone lesions, and soft tissues, with respective average accuracies of 19.9%, 43.3%, and 30.6%. This reduced performance in detecting these categories may impact the mAP50. However, this limitation is attributable to data constraints, as bone anomalies and soft tissues constitute only 0.41% and 0.68% of the total object dataset.

Figure 5 summarizes the overall detection performance of the iYOLOv8-AM model variants. For example, the iYOLOv8_GC variant achieved an average precision of 70% across all folds. In comparison, the fracture detection task achieved an average accuracy of 97.13%, significantly higher than the accuracy observed for the overall detection task. Figure Y illustrates this disparity, presenting performance metrics for both detection tasks across all model variants.

This study developed an improved version of the YOLOv8-AM model and introduced a novel variant incorporating the GC attention mechanism. These improved models were applied to detect pediatric DRF. Results showed that these improved models offer superior speed and accuracy in detecting DRF while reducing training time compared to the original versions. Integrating these improved models into clinical practice could provide significant benefits, including reduced costs and enhanced diagnostic efficiency.

Detecting fractures can vary significantly in complexity. While some fractures are straightforward to identify, minor fractures, particularly in the pediatric population, may be easily overlooked due to subtle signs on X-rays, especially near growth plates where fractures can be misleading [19]. Pediatricians undergo specialized training to recognize age-specific details, as accurate diagnosis and optimal treatment are essential to prevent growth disturbances [20]. However, the shortage of radiologists has increasingly limited access to pediatric radiologists, posing a potential threat to the quality of pediatric care, particularly in emergency settings [21].

In this context, AI and deep learning tools offer a promising solution. As the field of radiology evolves, AI is becoming an integral component in diagnostic imaging, enhancing accuracy and efficiency [22]. Research supports that AI-assisted diagnostics can significantly improve medical practice, leading to faster and more precise treatment decisions. For instance, Kwolek et al. demonstrated that AI tools could accurately assess Hallux Valgus severity and intermetatarsal ligaments compared to specialized physicians [23]. Similarly, Cohen et al. found that AI systems utilizing deep learning exhibited greater sensitivity than non-specialized radiologists in detecting wrist fractures [24]. Furthermore, it has been noted that emergency physicians have a sensitivity of 81.9% in detecting pediatric appendicular fractures [25].

The models' dataset and baseline parameters were based on the work of Chien et al. [16]. By modifying the training dataset, adjusting image sizes, and fine-tuning the model, notable improvements in accuracy were achieved. The newly introduced iYOLOv8-GC model attained a precision of 97.2%, surpassing many AI models reported in the literature. For instance, Binh et al.'s model achieved an accuracy of 88.2% for detecting radius fractures [26].

Results also highlight the advantages of deep learning in detecting DRF, particularly in terms of real-time detection speed. The improved model processes analyses in as little as 2.3 ms with the Resblock-GAM and at most 8.2 ms, outperforming both YOLOv8 and YOLOv9 models in speed. Additionally, the training time for the improved model is significantly shorter compared to the original models reported by Chien et al. [14, 16]

The proposed AI models utilizing deep learning are designed to complement, not replace, clinical expertise. They serve as efficient and accurate tools that enhance diagnostic processes. Numerous studies have demonstrated that integrating AI models can significantly improve physicians' accuracy in detecting fractures. For instance, access to model predictions increased the average resident's accuracy from 80–93% [27]. Similarly, combining AI models with non-specialized radiologists raised detection accuracy from 76–88% compared to non-specialized radiologists alone [24]. With reduced training and inference times, these models allow residents or junior physicians who may not specialize in pediatric radiology to integrate AI results with their clinical judgment for timely and accurate treatment decisions. The depository code and instructions on how to run the software of the iYOLO_v8_GC detecting tool are available online [28].

Future Applications

The 10-fold cross-validation results demonstrate the models’ reliability in generalizing across different subsets of data. The enhanced precision in fracture detection suggests that the model is particularly effective in identifying fractures, which could benefit clinical applications. Future research should explore additional features or model improvements to extend fracture detection capabilities in other body parts. It could be adapted to detect various types of fractures throughout the body with appropriate training and datasets (Fig. 6). The YOLO series' performance is highly dependent on the quality of the training data. As mentioned before, the accuracy for detecting bone anomalies, lesions, and soft tissues is limited by the current dataset's size diversity. The model would require additional training data specific to those fractures and the underrepresented classes to enhance accuracy for different fracture types. This extra training could enable the iYOLOv8-AM models to achieve reliable predictions with minimal data and training time.

The limitations of this study include the restricted scope of the dataset used for training the models, particularly the relatively small number of cases involving soft tissue, bone anomalies, and lesions. Additionally, there was a lack of parameterization to evaluate intermediate configurations, such as input image sizes of 720, 840, and 960. The study focused on YOLOv8; future research could explore similar improvements in YOLOv9 and subsequent models. Our work specifically targeted the development of the iYOLOv8-AM model for pediatric distal radius fractures, limiting its applicability to this area.

Ethics approval and consent to participate: Not applicable, since the study includes de-identified data available online.

Consent for publication: Not applicable, since the study includes de-identified data available online.

Availability of data and materials: the dataset analyzed during the current study are available in the data set/archive repository, https://figshare.com/articles/dataset/GRAZPEDWRI-DX/14825193

Competing Interests: The authors declare that they have no competing interests.

Funding: Not applicable

Authors’ Contributions: F.A. and A.P. trained and tested the model. M.M. wrote the introduction, results, and discussion, while A.P. and M.M. collaborated on writing the methodology. M.M. prepared the table and figures 2, 3, and 6, and A.P. prepared figures 1, 4, and 5. F.A. made the final edits to the manuscript, and all authors reviewed and approved the final version.

Acknowledgments: Not applicable

Hedström EM, Svensson O, Bergström U, Michno P. Epidemiology of fractures in children and adolescents. Acta Orthop. 2010;81:148–53.
Miele V, Galluzzo M, Trinci M. Missed Fractures in the Emergency Department. In: Romano L, Pinto A, editors. Errors in Radiology. Milano: Springer Milan; 2012. pp. 39–50.
Rimmer A. Radiologist shortage leaves patient care at risk, warns royal college. BMJ. 2017;359:j4683.
Berlin L. Defending the missed radiographic diagnosis. AJR Am J Roentgenol. 2001;176:317–22.
Mounts J, Clingenpeel J, McGuire E, Byers E, Kireeva Y. Most frequently missed fractures in the emergency department. Clin Pediatr (Phila). 2011;50:183–6.
Malamateniou C, Knapp KM, Pergola M, Woznitza N, Hardy M. Artificial intelligence in radiography: Where are we now and what does the future hold? Radiogr Lond Engl 1995. 2021;27(Suppl 1):S58–62.
Patel BN, Rosenberg L, Willcox G, Baltaxe D, Lyons M, Irvin J, et al. Human-machine partnership with artificial intelligence for chest radiograph diagnosis. NPJ Digit Med. 2019;2:111.
Blüthgen C, Becker AS, Vittoria de Martini I, Meier A, Martini K, Frauenfelder T. Detection and localization of distal radius fractures: Deep learning system versus radiologists. Eur J Radiol. 2020;126:108925.
Gan K, Xu D, Lin Y, Shen Y, Zhang T, Hu K, et al. Artificial intelligence detection of distal radius fractures: a comparison between the convolutional neural network and professional assessments. Acta Orthop. 2019;90:394–400.
Currie G, Hawk KE, Rohren E, Vial A, Klein R. Machine Learning and Deep Learning in Medical Imaging: Intelligent Imaging. J Med Imaging Radiat Sci. 2019;50:477–87.
Kim J, Lee S, Hwang E, Ryu KS, Jeong H, Lee JW, et al. Limitations of Deep Learning Attention Mechanisms in Clinical Research: Empirical Case Study Based on the Korean Diabetic Disease Setting. J Med Internet Res. 2020;22:e18418.
Lindsey R, Daluiski A, Chopra S, Lachapelle A, Mozer M, Sicular S, et al. Deep neural network improves fracture detection by clinicians. Proc Natl Acad Sci U S A. 2018;115:11591–6.
Niu Z, Zhong G, Yu H. A review on the attention mechanism of deep learning. Neurocomputing. 2021;452:48–62.
Chien C-T, Ju R-Y, Chou K-Y, Chiang J-S. YOLOv9 for Fracture Detection in Pediatric Wrist Trauma X-ray Images. 2024.
Ju R-Y, Chien C-T, Lin C-M, Chiang J-S. Global Context Modeling in YOLOv8 for Pediatric Wrist Fracture Detection. 2024.
Chien C-T, Ju R-Y, Chou K-Y, Xieerke E, Chiang J-S. YOLOv8-AM: YOLOv8 with Attention Mechanisms for Pediatric Wrist Fracture Detection. 2024.
Global Context Networks (GCNet). Explained | Paperspace Blog. Paperspace by DigitalOcean Blog. 2021. https://blog.paperspace.com/global-context-networks-gcnet/. Accessed 23 Sep 2024.
Guo M-H, Xu T-X, Liu J-J, Liu Z-N, Jiang P-T, Mu T-J, et al. Attention mechanisms in computer vision: A survey. Comput Vis Media. 2022;8:331–68.
George MP, Bixby S. Frequently Missed Fractures in Pediatric Trauma: A Pictorial Review of Plain Film Radiography. Radiol Clin North Am. 2019;57:843–55.
Liao JCY, Chong AKS. Pediatric Hand and Wrist Fractures. Clin Plast Surg. 2019;46:425–36.
Farmakis SG, Chertoff JD, Barth RA. Pediatric Radiologist Workforce Shortage: Action Steps to Resolve. J Am Coll Radiol JACR. 2021;18:1675–7.
Dundamadappa SK. AI tools in Emergency Radiology reading room: a new era of Radiology. Emerg Radiol. 2023;30:647–57.
Kwolek K, Gądek A, Kwolek K, Kolecki R, Liszka H. Automated decision support for Hallux Valgus treatment options using anteroposterior foot radiographs. World J Orthop. 2023;14:800–12.
Cohen M, Puntonet J, Sanchez J, Kierszbaum E, Crema M, Soyer P, et al. Artificial intelligence vs. radiologist: accuracy of wrist fracture detection on radiographs. Eur Radiol. 2023;33:3974–83.
Gasmi I, Calinghen A, Parienti J-J, Belloy F, Fohlen A, Pelage J-P. Comparison of diagnostic performance of a deep learning algorithm, emergency physicians, junior radiologists and senior radiologists in the detection of appendicular fractures in children. Pediatr Radiol. 2023;53:1675–84.
Binh LN, Nhu NT, Vy VPT, Son DLH, Hung TNK, Bach N, et al. Multi-Class Deep Learning Model for Detecting Pediatric Distal Forearm Fractures Based on the AO/OTA Classification. J Imaging Inf Med. 2024;37:725–33.
Zech JR, Carotenuto G, Igbinoba Z, Tran CV, Insley E, Baccarella A, et al. Detecting pediatric wrist fractures using deep-learning-based object detection. Pediatr Radiol. 2023;53:1125–34.
Mathew A. Aashikmathewcodes/Improved-iYOLO-v8-Fracture-detection-for-pediatric-wrist. 2024.

No competing interests reported.

Enhancing Pediatric Distal Radius Fracture Detection: Optimizing YOLOv8 with Advanced AI and Machine Learning Techniques.

Status:

Version 1

Abstract

Figures

Introduction

Methodology

=(F) (F)

Results

Discussion

Future Applications

Declarations

References

Additional Declarations

Status:

Version 1